"Smart people learn from their mistakes. Geniuses learn from GitHub issues."
Imagine an AI that doesn’t just write perfect code — but actually studies thousands of real-world bugs, failures, and pull request debates to understand how things go wrong.
That’s what I tried to build.
This is the story of how I scraped, trained, and deployed a local LLM that doesn’t just generate code — it warns me about bugs real devs have already made. And it’s powered entirely by GitHub’s open issue tracker.
🔥 The Problem: Tutorials Don’t Teach You What Breaks
As developers, we spend hours on tutorials that show us:
- “How to set up a REST API”
- “How to train a basic LLM”
- “How to deploy a project with Docker”
But these are happy path instructions. They show you the "golden path."
Meanwhile, on GitHub:
- Issues filled with edge cases
- Pull request discussions full of trade-offs
- "Why was this line changed?" mysteries
That’s the real learning — the dark matter of dev education.
So I asked myself:
Can I make an AI that doesn’t learn from code examples, but from code mistakes?
🧠 The Idea: Train an LLM on GitHub Issues & Fixes
Instead of feeding an AI perfect Stack Overflow answers, I did this:
- Crawled open-source repos with high issue + PR activity
- Extracted:
- Title + body of bug reports
- Linked PRs that fixed them
- Reviewer comments
- Chunked each (issue → fix) pair into embeddings
- Indexed them in a vector DB
- Created a CLI and VS Code extension where I could ask:
“Has anyone fixed a bug like this before?”
And shockingly... it worked.
🛠️ My Stack: Building the "Bug Sage"
| Component | Tool Used |
|---|---|
| Issue Scraping | GitHub API + GraphQL |
| Embedding |
text-embedding-ada-002 via OpenAI OR Instructor-XL locally |
| Vector DB | ChromaDB |
| Retrieval | LangChain |
| UI | CLI + VS Code Sidebar |
| Local LLM |
Phi-3 or Mistral via Ollama |
🐛 How It Works (Real Example)
Say I’m debugging a KeyError in my FastAPI app when deploying to AWS Lambda.
Instead of googling aimlessly or hitting Stack Overflow, I type:
bugsage "KeyError during AWS Lambda cold start in FastAPI app"
And it retrieves this issue from another repo:
📝
#328 - FastAPI app fails on cold start due to environment variables missing
🔧 Fixed by moving.envloading inside the handler in PR #329
Suddenly, I’m not just getting a fix.
I’m getting context, explanation, and real-world patterns.
🤖 The Coolest Part? The AI Learns With Every Crawl
Every weekend, a GitHub Action re-scrapes:
- New issues from starred repos
- PRs with
fixeskeywords - Tags like
bug,regression,performance
It self-indexes all this into ChromaDB. Over time, the "Bug Sage" gets smarter — like a developer mentor who reads every project on GitHub for you.
- Check this out, if u have some moment of time: (while reading it)
🧠 The Educational Value: Learning From Pain Points
This isn’t just a productivity tool.
It’s a learning engine.
You start to see:
- How real teams debug
- What kinds of mistakes repeat
- Why certain decisions are controversial
You start coding defensively — with foresight.
It’s like pair programming with 1,000 senior devs whispering, “Hey… that didn’t work for us either.”
🤯 Unexpected Use Cases
- ✅ Code review help: It suggests real PR debates for similar changes
- 📉 Prevent regressions: Matches code diffs to past rollback issues
- 🎓 Learning prompts: “Give me 3 bugs people faced with WebSockets + Django Channels”
- 🕵️♂️ Open-source archaeology: “What were the most common bugs in X repo over 2 years?”
😂 Dev Humor (You Know It’s Coming)
- 🧟 “I don’t make the same bug twice… I make it 10 times, slightly differently”
- 🧙♂️ “Bug Sage, what’s the ancient wisdom on async DB calls?”
- 👶 Me: "Why is my code crashing?" GitHub AI: "It has happened before… and it will happen again."
📚 How You Can Build Your Own “Bug Sage”
Want to try this at home?
Step 1: Crawl GitHub issues and PRs
from github import Github
g = Github("your_token")
repo = g.get_repo("tiangolo/fastapi")
issues = repo.get_issues(state="closed", labels=["bug"])
Step 2: Pair issues to PRs
Look for text like "Fixes #123" in PR bodies.
Step 3: Embed text
Use:
from langchain.embeddings import OpenAIEmbeddings
Or local models like Instructor-XL via HuggingFace.
Step 4: Store + Query with Chroma
from langchain.vectorstores import Chroma
Step 5: Build CLI or integrate into VS Code!
🌍 Final Thought: Make GitHub Your Mistake Mentor
We often treat GitHub as a place to show perfect work.
But it’s really a museum of broken code — and if you mine it well, you’ll learn ten times faster than any tutorial.
Don’t just write code.
Study how it breaks — and let AI help you never repeat it.
💬 Tired of Building for Likes Instead of Income?
I was too. So I started creating simple digital tools and kits that actually make money — without needing a big audience, fancy code, or endless hustle.
🔓 Premium Bundles for Devs. Who Want to Break Free
These are shortcuts to doing your own thing and making it pay:
🌍 I built a simple website for a local biz and got $500+ — No design skills. Just solved a real problem.
🚀 Launched a SaaS in 7 days — no code, no audience — It’s messy but it works.
🔌 Used public APIs to build tiny tools people paid $997 for — Took what was already out there and made it useful.
📦 $300 in 3 days from a simple resource vault — Just organized links + tools. That’s it.
📈 Ranked a local site without writing a single blog post — SEO doesn’t have to be hard if you do it differently.
🔧 Quick Kits (Take 1 Product That Actually Works for You)
These are personal wins turned into plug-and-play kits — short instruction guides:
⚡ $1K in a week using APIs I didn’t even build — Copy-paste logic, add polish, publish.
🔥 My $0 dev setup now earns $97+ daily — Took years to build. Now it runs quietly in the background.
💼 This SaaS starter kit sells itself for $499 — Turns out, people love skipping setup pain.
📚 I turned academic papers into real products — It’s all just buried gold if you know where to look.
💡 My dev portfolio became a $297 product — I just told my story and sold the assets I made along the way.
👉 Browse all tools and micro-business kits here
👉 Browse all blueprints here
Top comments (0)